Re: [PATCH v2 09/22] sched/fair: Use IPC class score to select a busiest runqueue

[Ricardo Neri <ricardo.neri-calderon@linux.intel.com>]

On Wed, Dec 14, 2022 at 11:16:39PM +0000, Ionela Voinescu wrote:
> Hi Ricardo,
> 
> On Tuesday 13 Dec 2022 at 16:32:43 (-0800), Ricardo Neri wrote:
> [..]
> > > >  /**
> > > > @@ -10419,8 +10442,8 @@ static struct rq *find_busiest_queue(struct lb_env *env,
> > > >  {
> > > >  	struct rq *busiest = NULL, *rq;
> > > >  	unsigned long busiest_util = 0, busiest_load = 0, busiest_capacity = 1;
> > > > +	int i, busiest_ipcc_delta = INT_MIN;
> > > >  	unsigned int busiest_nr = 0;
> > > > -	int i;
> > > >  
> > > >  	for_each_cpu_and(i, sched_group_span(group), env->cpus) {
> > > >  		unsigned long capacity, load, util;
> > > > @@ -10526,8 +10549,37 @@ static struct rq *find_busiest_queue(struct lb_env *env,
> > > >  
> > > >  		case migrate_task:
> > > >  			if (busiest_nr < nr_running) {
> > > > +				struct task_struct *curr;
> > > > +
> > > >  				busiest_nr = nr_running;
> > > >  				busiest = rq;
> > > > +
> > > > +				/*
> > > > +				 * Remember the IPC score delta of busiest::curr.
> > > > +				 * We may need it to break a tie with other queues
> > > > +				 * with equal nr_running.
> > > > +				 */
> > > > +				curr = rcu_dereference(busiest->curr);
> > > > +				busiest_ipcc_delta = ipcc_score_delta(curr,
> > > > +								      env->dst_cpu);
> > > > +			/*
> > > > +			 * If rq and busiest have the same number of running
> > > > +			 * tasks, pick rq if doing so would give rq::curr a
> > > > +			 * bigger IPC boost on dst_cpu.
> > > > +			 */
> > > > +			} else if (sched_ipcc_enabled() &&
> > > > +				   busiest_nr == nr_running) {
> > > > +				struct task_struct *curr;
> > > > +				int delta;
> > > > +
> > > > +				curr = rcu_dereference(rq->curr);
> > > > +				delta = ipcc_score_delta(curr, env->dst_cpu);
> > > > +
> > > > +				if (busiest_ipcc_delta < delta) {
> > > > +					busiest_ipcc_delta = delta;
> > > > +					busiest_nr = nr_running;
> > > > +					busiest = rq;
> > > > +				}
> > > >  			}
> > > >  			break;
> > > >  
> > > 
> > > While in the commit message you describe this as breaking a tie for
> > > asym_packing,
> > 
> > Are you referring to the overall series or this specific patch? I checked
> > commit message and I do not see references to asym_packing.
> 
> Sorry, my bad, I was thinking about the cover letter, not the commit
> message. It's under "+++ Balancing load using classes of tasks. Theory
> of operation".
> 
> > 
> > > the code here does not only affect asym_packing. If
> > > another architecture would have sched_ipcc_enabled() it would use this
> > > as generic policy, and that might not be desired.
> > 
> > Indeed, the patchset implements support to use IPCC classes for asym_packing,
> > but it is not limited to it.
> > 
> 
> So is your current intention to support IPC classes only for asym_packing
> for now?

My intention is to introduce IPC classes in general and make it available
to other policies or architectures. I use asym_packing as use case.


> What would be the impact on you if you were to limit the
> functionality in this patch to asym_packing only?

There would not be any adverse impact.

> 
> > It is true that I don't check here for asym_packing, but it should not be a
> > problem, IMO. I compare two runqueues with equal nr_running, either runqueue
> > is a good choice. This tie breaker is an overall improvement, no?
> > 
> 
> It could be, but equally there could be other better policies as well -
> other ways to consider IPC class information to break the tie.
> 
> If other architectures start having sched_ipcc_enabled() they would
> automatically use the policy you've decided on here. If other policies
> are better for those architectures this generic policy would be difficult
> to modify to ensure there are no regressions for all other architectures
> that use it, or it would be difficult to work around it.
> 
> For this and for future support of IPC classes I am just wondering if we
> can better design how we enable different architectures to have different
> policies.

I see your point. I agree that other architectures may want to implement
policies differently. I'll add an extra check for env->sd & SD_ASYM_PACKING.

Thanks and BR,
Ricardo